llm-ingestion-okf/docs/2026-09-08-k3-runde5-hitat8-og-skriftakse.md

K3 round 5: hit@8 on both K2 bundles, and typography as a reserve

Date: 2026-09-08 · HEAD at start: 53d5c74 · Suite at start: 1441

Round 4 shipped two PDF readers and left three questions behind it. This round answers the retrieval one first, because it is the one that could have reversed a default, and then measures the segmentation candidate the round-4 report pointed at. Two of the three answers are negative, and the negative ones are the informative ones.

0. What this round found, in one paragraph

The default move did not cost anything on the retrieval side: hit@8 is 5 of 6 on both K2 bundles, with the same five ranks at 1 and the same row missing on both. The decision rule therefore does not fire and --unit-fold stays in the default. The font reserve — typography read only where the outline arm finds no run — was built, is off, and must stay off: it cannot reach the position it was designed for, because that position numbers its own chapters. And the xlsx re-reading works, but only if BOTH halves of it move together; the number alone makes the cell worse.

Two published figures are corrected here rather than left standing. Both are this repository's own.

---

1. Reproduction of the starting state

Every figure below was measured on 53d5c74 with a clean tree.

what	expected	measured
git rev-parse HEAD	53d5c74	53d5c74c96b4316dcb3380227d72be908e1be270
git status --short	empty	empty
pytest -q	1441	1441 passed
ruff check src tests tools	clean	All checks passed!
mypy src	clean	Success: no issues found in 21 source files
okf build <folder>, no flags	30 md	30, exit 0, 0/5 rejected
same, --keep-table-heading --sheet-section-rows --drop-wrapped-outline	52 md	52, exit 0, 0/5 rejected
unpushed commits	4	4 (47f6619, efa92ce, f6fea13, 53d5c74)
tag v0.6.0	on f6fea13	on f6fea13, one commit behind HEAD

Both cached bundles were counted before use rather than trusted: the Arm B tree holds 1108 md files and round 4's default artifact 944, so neither was rebuilt.

---

2. hit@8 on both K2 bundles

One instrument (tools/okf_consume_measure.py), one gold set, one --k 8, run against both bundles. The gold set is local-only and neither the questions nor the gold documents appear here; the row number carries the identity, as in docs/2026-09-07-okf-konsumskill-maaling.md.

Bundle 1 (Arm B): 629 concepts, 1108 files, every arm flag off. Bundle 2 (default): 492 concepts, 944 files, digest bdefa679957edd7d…, default flags and default epoch stamp.

Both bundles are the same 39-document readable set, so documents is 39 in both and the chance baselines are directly comparable.

row	gold concepts (B)	hit (B)	rank (B)	considered (B)	spent (B)	chance (B)	gold concepts (D)	hit (D)	rank (D)	considered (D)	spent (D)	chance (D)
1	1	yes	1	629	76 897	0.013	1	yes	1	492	43 288	0.016
2	49	yes	1	629	21 403	0.479	26	yes	1	492	18 458	0.354
3	20	yes	1	629	26 282	0.229	17	yes	1	492	28 347	0.247
4	43	yes	1	629	19 818	0.434	36	yes	1	492	21 125	0.458
5	11	yes	1	629	22 459	0.132	18	yes	1	492	17 866	0.259
6	5	no	–	629	20 282	0.062	5	no	–	492	33 113	0.079
total		5 of 6				1.35 expected by chance		5 of 6				1.41 expected by chance

0 of 6 rows lost. The decision rule in the order fires at >= 2 of 6, so it does not fire: --unit-fold stays in the default, and no DEFAULT_UNIT_FOLD was touched. This reproduces the PM measurement of 21:20Z exactly, including the ranks.

The gold sets change size while the rank does not. 49→26, 20→17, 43→36 and 11→18 concepts between the two bundles, and every one of those rows still ranks

1. That is the fold merging concepts rather than removing a document from the top — which is the property the fold was defended on, measured here on the retrieval side for the first time.

Both known-negative controls behave identically on both bundles: 2 of 6 tokens reaching nothing on the planned negative, 4 of 6 on the lexically disjoint one, 8 excerpts delivered in each case. The negatives are a control on the question, not on the ranker, and they did not move.

2.1 The S7 control row, and a correction to how it was measured

A separate question with a separate answer, measured with --cost-vocabulary --k 12 on both bundles:

	Arm B (629)	default (492)
priced concept present	yes	yes
candidate rank (whole ranking)	10 of 629	19 of 492
rank among concepts that survive the gate	10 of 280	19 of 280
delivered at --k 12	no	no
why not	over_budget_after_knapsack — inside the shortlist, evicted; the excerpt weighs 68 046 of a 120 000 budget	below_k — outside the shortlist at k = 12
excerpts delivered	11	12
budget spent	84 953	51 810

Two published numbers are corrected by this row.

First, the ranks. Round 3 and round 4 published this concept at 96 of 629 and 159 of 492. Both are reproducible and both name a mixed configuration: the measuring script (/tmp/k3r4/rank4.py) passes cost_vocabulary=True to concept_scores but calls document_scores(root, question) with the vocabulary left off, while build_payload passes it to both. Scored the way the shipped payload scores it, the same concept is 10 of 629 and 19 of 492. The published figures are not wrong about their own arithmetic; they are answers about a ranking the library never emits.

Second, the attribution. Round 4 wrote that the non-delivery "is the default move of 47f6619/f6fea13 showing up on the retrieval side for the first time". Measured here on the Arm B bundle with today's code, the concept is not delivered there either. The default move is not the cause; the two bundles fail to deliver it for two different reasons, and only one of them is about bundle size. The knapsack eviction on Arm B is the case --reserve-top-rank was built for and cannot fix, because the reservation is for rank one and this is rank ten — which is exactly what the flag's own documentation already says.

What round 4 could not have known is that the Arm B column was never measured; it was carried over from round 3's own build. Re-running it was the whole correction.

---

3. The font reserve

Round 4 measured the font reader two ways and shipped neither: on TOP of Arm D it took pdf from 2 of 8 to 0 of 8, and INSTEAD of Arm D (--outline-run 0) it scored 1 of 8 while landing position 7 on 30 = its reference count. The hypothesis this round was asked to test: use typography only where Arm D's outline gate admits no run, and position 7 is bought without paying positions 2 and 8.

The hypothesis is falsified, and it is falsified by the condition itself rather than by a score.

3.1 Where the reserve can fire, on the 12-position reference

Measured before any code was written, on the cached round-3 rendering:

position	type	outline runs at minimum 3	reserve
0	pdf	0	fires
1	pdf	2 (last run 7 long)	silent
2	pdf	2 (last run 9)	silent
3	xlsx	0	fires (no effect: the font reader is PDF-only)
4	pdf	1 (last run 5)	silent
5	docx	0	fires (no effect: PDF-only)
6	pdf	1 (last run 3)	silent
7	pdf	3 (last run 10)	silent
8	pdf	2 (last run 8)	silent
9	pdf	3 (last run 6)	silent
10	docx	0	fires (no effect: PDF-only)
11	docx	0	fires (no effect: PDF-only)

Position 7 — the one position the flag exists for — has three outline runs, the last of them ten entries long. No reading of "no run at all" reaches it: it is a run at minimum 3, at minimum 1, and at every minimum up to 10. The 1 of 8 that round 4 measured under --outline-run 0 came from removing Arm D on that document, not from adding a second source where there was none.

3.2 What it scores

pos	type	ref |F|	default	+ reserve	+ font everywhere
0	pdf	undefined	for grovt (U1)	for grovt (U1)	for grovt (U1)
1	pdf	15	for fint (U17, d2)	for fint (U17, d2)	for grovt (U5, d10)
2	pdf	9	treffer (d0)	treffer (d0)	for fint (U12, d3)
3	xlsx	11	for grovt (U0, S1, d12)	for grovt (U0, S1, d12)	for grovt (U0, S1, d12)
4	pdf	1	for fint (U4, d3)	for fint (U4, d3)	for fint (U11, d10)
5	docx	3	treffer (d0)	treffer (d0)	treffer (d0)
6	pdf	6	for fint (U7, d1)	for fint (U7, d1)	for fint (U11, d5)
7	pdf	30	for fint (U33, d3)	for fint (U33, d3)	for fint (U33, d3)
8	pdf	8	treffer (d0)	treffer (d0)	for fint (U16, d8)
9	pdf	4	for fint (U10, d6)	for fint (U10, d6)	for fint (U12, d8)
10	docx	2	treffer (d0)	treffer (d0)	treffer (d0)
11	docx	1	treffer (d0)	treffer (d0)	treffer (d0)
			5 of 12 · pdf 2 of 8 · docx 3 of 3 · xlsx 0 of 1	5 of 12, identical cell for cell	4 of 12 · pdf 1 of 8

The reserve changes not one cell of the twelve. It fires on five positions and alters none of them: position 0's glyphs carry no ToUnicode mapping, so its font rendering is byte-identical to its plain one, and the other four are office documents the PDF reader never touches. The default column reproduces round 4 exactly (pdf 2 of 8, docx 3 of 3, xlsx 0 of 1), which is what makes the instrument calibrated rather than merely re-run.

The acceptance criterion the order set — position 7 hits 30 without losing 2, 6 and 8 — is unreachable by construction, not merely unmet. And position 7's (c) in the reference is I, so the criterion's literal gate forbids a hit there in any case; both readings stood in round 4 § 3 and both stand here.

3.3 What the reserve reaches on the whole corpus

The twelve positions are a sample of 43. Over the full cached corpus:

	count
readable documents	39
of which pdf	32
pdf admitting no outline run at minimum 3	10
of those, whose font rendering differs at all	4
of those, whose plan changes	4 (1→4, 1→2, 0→2 and 5→6 entries)

So the flag is not inert: it acts on 4 of 39 readable documents. But none of those four is in the rated sample, so there is no reference count for any of them, and this round cannot say whether the four changes are improvements. A flag with a reach and no measurable quality signal is exactly a flag that stays off.

3.4 What was built anyway, and why

--pdf-headings font-reserve: a third value on the option round 4 introduced, so none, font and font-reserve are three answers to one question and no caller can ask for two at once.

The condition lives in one function, propose.heading_reserve_applies(text, *, outline_run), and the proposer and the door both call it. That is not tidiness: a plan indexes the exact string it was proposed against (text_sha256), so a reserve that fired on one side and not the other would turn every document it touches into a coded rejection. The door receives it as a predicate (heading_reserve: Callable[[str], bool] | None), the same shape gate already has, because the condition is the proposer's outline grammar and Door B does not own that grammar — a direct import would point the dependency against the layering.

The reserve reads the gate as configured. At --outline-run 0 the gate admits nothing at all, so the reserve becomes unconditional; that combination is round 4's "font instead of Arm D" at 1 of 8, and a caller reaching it does so deliberately. This is stated in the code and pinned by a test.

Control: the default build of the operator's folder is byte-identical before and after the change (diff -r, 30 md files either way), and the suite went 1441 → 1449 with eight new tests, three of which were red first for three different reasons (missing predicate, wrong expected segmentation, wrong report field).

---

4. The xlsx reference count, re-read as 12

The operator's re-reading is that the priced worksheet holds 11 numbered cost groups plus one leading block of real prose that has to live somewhere, so 11 + 1 = 12. Measured on the extraction, under --sheet-section-rows:

unit	rule	count
the leading block	rule:table-block	1
the cost groups	rule:sheet-section	11
total		12

The rule produces exactly 12 entries, and the split is 11 + 1 in the shape the re-reading describes. The re-reading is confirmed on the artifact.

But the number alone does not deliver the hit, and this is the finding. The criterion as written treats a table-derived concept under b = M as S — a table that should have been merged into a heading — not as a unit. So:

reading of the reference	U	T	S	distance	label
|F| = 11, leading block counted as S (rounds 3 and 4)	11	0	1	1	for fint
|F| = 12, leading block still counted as S	11	0	1	2	for grovt
|F| = 12, leading block counted as the 12th unit	12	0	0	0	treffer

Changing the number without changing what the criterion counts makes the cell worse, from distance 1 to distance 2. xlsx reaches 1 of 1 only when both halves move together.

This is the operator's reference, not mine. What is measured here is that the artifact contains 11 + 1 units in exactly the shape the re-reading names, and that a hit requires ratifying the second half too. Rounds 3 and 4 both used the reading "11" and both reported distance 1, so nothing published so far depends on the outcome either way.

---

5. --sheet-section-rows as a default: what the 12 positions say

Because § 4 raised it, the D1/D3 configuration was scored against the shipped default on all twelve positions:

pos	default	+ --keep-table-heading --sheet-section-rows --drop-wrapped-outline
3 (xlsx)	for grovt, d12	for fint, d1 (treffer, d0 under the re-read reference)
4 (pdf)	for fint, d3	treffer, d0
9 (pdf)	for fint, d6	for fint, d2
all other nine	—	identical, cell for cell

Three positions improve, none regresses, taking pdf from 2 of 8 to 3 of 8 and leaving docx at 3 of 3. That is the strongest segmentation result this round has, and it is not acted on here: the order makes a default move conditional on the K2 control as well, and it is a fourth change to every consumer's bytes in two days. The K2 control was run for exactly that reason and it does not hold (§ 7), so the default stays where it is.

---

6. What still stands, unchanged

• The prefix over-matches the other way (round 3 finding, untouched since): a compound word matches a four-character prefix, df 159 of 270. It belongs to the rarity weight and was not this round's work.
• pdf position 1 (17 against a reference of 15) is now refused on four axes: number, outline, typography, and the reserve — the reserve is silent there, because that document has two outline runs.
• vegnormal-okf FUNN 1 and FUNN 2 are still not investigated in this repo.
• tests/test_okf_consume.py:940 pins hit@8 against the Arm B bundle alone. The default bundle is pinned nowhere. That is a gap, and it is named in STATE's next block rather than closed here: O4/O5 fence that file, and this round only ran the instrument.

---

7. The K2 control for the D1/D3 configuration: it moves, so nothing moves

The whole 43-document corpus rebuilt with --keep-table-heading --sheet-section-rows --drop-wrapped-outline on top of the shipped default. Exit 0, 2305 s wall (773 s inside the door), 39/43 persisted, 4/43 coded rejections -- the same conservation identity as the default build.

	default	+ D1/D3
concepts reachable from the index tree	492	491
md files	944	942
hit@8 over the six questions	5 of 6	5 of 6
expected by chance	1.41	1.57

row	gold concepts (default)	rank	gold concepts (D1/D3)	rank
1	1	1	12	2
2	26	1	26	1
3	17	1	17	1
4	36	1	36	1
5	18	1	17	1
6	5	--	5	--

The count holds and the control does not. Row 1's gold document is split from one concept into twelve by the sheet-section rule, and its best concept then ranks 2 instead of 1. It is still a hit at 8, and it is still a change to a published number on a corpus the operator reads.

The order's condition for a default move is "no regression on the 43-document corpus or the K2 control". The 12-position reference improves three cells and regresses none (§ 5); the K2 control regresses one row's rank. The condition is therefore not met and the default is not moved. The flags stay reachable and off, exactly as they were.

That is worth stating plainly rather than as a near miss: the segmentation side and the retrieval side disagreed about this configuration, and the disagreement is real -- splitting a worksheet into its rows is what the reference asks for and what costs the ranker a rank position, because twelve small concepts each carry less of the question than one whole sheet did.

---

8. Honesty limits

• n = 6 questions on the retrieval side, and n = 12 positions on the segmentation side, of which 8 pdf, 3 docx and 1 xlsx. Every per-type figure in this report is a fraction of 8, 3 or 1. A "3 of 8" is three documents.
• One rater, one reference. The twelve-cell reference and the criterion that scores it were authored by the same operator, and neither has an independent second rating. Where the criterion and the reference disagree (§ 4), that is visible only because the disagreement was arithmetic.
• Both bundles are the same 43-document corpus, so "hit@8 held on both" is a statement about one corpus read two ways, not about two corpora.
• "Delivered" is not "answered". hit@8 says a gold document appeared among eight excerpts. Nothing here measures whether a reader could answer the question from those excerpts; the one measurement that ever tried (docs/2026-09-07-okf-konsumskill-maaling.md) got the correct answer on 1 of 3 bundles.
• The reserve's corpus reach (§ 3.3) was measured by simulation, not by running okf build --pdf-headings font-reserve over all 43 documents: the two cached renderings were combined by the same predicate the shipped code calls. The predicate is identical; the extraction path around it was exercised only on the fixtures and on the 12-document sample.
• The four documents the reserve changes have no reference count. They are outside the rated sample, so "4 of 39 change" is a reach, not a quality claim.
• The corrected S7 ranks (§ 2.1) were measured on today's code. Round 3's own build is not available, so this report cannot separate "the instrument was mixed" from "the code also moved" for the delivery outcome; it can and does separate it for the rank, because the rank reproduces exactly under the published instrument and changes only when the vocabulary is passed consistently.